Reflections on Safety and Artificial Intelligenceerichorvitz.com/OSTP-CMU_AI_Safety_framing_talk.pdfEric Horvitz Reflections on Safety and Artificial Intelligence Pittsburgh, PA June

Eric Horvitz

Reflections on Safety

and Artificial Intelligence

Pittsburgh, PAJune 27, 2016

Exploratory Technical Workshop on Safety and Control for AICarnegie Mellon University

AI & Safety

Constellation of methods referred to as Artificial Intelligence will touch our lives more closely and intimately

AI moving into high-stakes applicationsHealthcareTransportationFinancePublic policyDefense

Much to do on principles, methods, and best practices

Safe AI Systems

Relevance of Multiple Subdisciplines

Safe AI Systems


MultiagentSystems

Mechanism Design

Planning

SensorFusion

ControlTheory

Machine Learning

Metareasoning

RobustOptimization

HCI

Verification

Security

Safe AI Systems


MultiagentSystems

Mechanism Design

Planning

SensorFusion

ControlTheory

Machine Learning

Metareasoning

RobustOptimization

HCI

Verification

Security

Safety-Critical Systems

safetyˈsāftē/

noun1. the condition of being protected from or unlikely to cause danger, risk, or injury

safety-critical ˈsāftēˌkridək(ə)l/

adjective1. systems whose failure could result in loss of life, significant property damage, or damage to the environment.

2. designed or needing to be fail-safe for safety purposes.

adjectiveincorporating some feature for automatically counteracting the effect of an anticipated possible source of failure

noundevice or practice that, in the event of a failure, responds or results in a way that will cause no harm, or at least minimizes harm.

fail-safe \ˈfāl-ˌsāf\

Fail-safe

George Westinghouse, 1869Train braking system

Fail-safe designAir brakes

Brakes held "off" actively by healthy system

Brakes naturally resort to “on” if any failure of braking system

Fail-safe




Fail-safe practiceFull-power throttle on arrested landing

Fail-safe




Fail-safe planFree return trajectory

Fail-safe


✓ Mechanism✓ Practice✓ Plan



Fail-safe planFree return trajectory

Fail-safe


✓ Mechanism✓ Practice✓ Plan

✓ Monitoring



Growing interest in issues & directions with AI in real-world settings

Grappling with uncertainty and more general incompleteness

AAAI President’s address (2008), “Artificial Intelligence in the Open World.”

AAAI President’s address (2016), “Steps Toward Robust Artificial Intelligence.”

AI in the Open World

E. Horvitz. Artificial Intelligence in the Open World, AAAI President’s Address, Chicago, IL, July 2008.

T. Dietterich, Steps Toward Robust Artificial Intelligences, AAAI President's Address, Phoenix, AX. February, 2016.

http://research.microsoft.com/en-us/um/people/horvitz/AAAI_Presidential_Lecture_Eric_Horvitz.htm

http://web.engr.oregonstate.edu/~tgd/talks/dietterich-aaai-presidents-address-final.pdf

Open-world complexity incomplete understanding

Uncertainties & poor-characterization of performancePoor operating regimes, unfamiliar situations

Challenges of transfer across time & space

Special Considerations with AI

Rich ontology of failures

Numerous failure modalities New attack surfaces (e.g., machine learning attack)Self-modification & gaming (e.g., modify reward fcn)Unmodeled influences

Challenge of coordinating human-machine collaborations

Operational opacity

AI & Open-World Complexity

Frame problem

How to tractably derive consequences of an action?

Qualification problem

Understanding preconditions required for actions to have intended effects

Ramification problem

Understanding all important effects of action


Rise of probabilistic methods: known unknowns

Recent attention to unknown unknowns




Decision making under uncertainty & incompleteness




























Fang, et al., 2015

Direction: Learn about abilities & failures

Performance

Successes & failures

p( fail | E, t)

Predictive model of

confidenceImage

H1

H2

H3

W1

W2

W3

W4

Input s

H3

Caption:

a man holding a tennis

racquet on a tennis court

H1

H2

H3

W1

W2

W3

Input t1

H3

W4

Deep learning about deep learning performance

InferredState

InferenceReliability

Inference Inference

Performance


p( fail | E, t)

Predictive model of

confidence


Toyama & H. 2000

http://research.microsoft.com/~horvitz/modfusion.pdf

InferredState

InferenceReliability

Inference Inference

Performance


p( fail | E, t)

Predictive model of

confidence


Toyama & H. 2000


Inference modality A

Inference modality B

Inference modality C

Direction: Robustness via analytical portfolios

Toyama & H. 2000



Toyama & H. 2000



Toyama & H. 2000

lights out

jolted camera

facing away

periph. distraction

back. subtract color based motion decay

Perceptual modalitiesUnmodeled situations in open world

Joint inference


Vary model structure, parameters, inferences

Direction: Understanding robustness via sensitivity analyses

Vary model structure, parameters, inferences

Direction: Understanding robustness via sensitivity analyses

Direction: Robust optimization to minimize downside

Tamar, 2015; Chow, et al., 2014; per Dietterich, AAAI lect. 2016

Robust optimization under uncertain parameters

Risk-sensitive objective

e.g., conditional-value-at-risk budget

Methods trade upside value

for reducing probability of

costly outcomes

Direction: Learn about unknown unknowns

Data, experience, rich simulations

Detect anomalies, unexpected variations, distributional shifts

Meta-analysis & transfer

Human engagement


“Beat the Machine” (Attenberg, Ipeirotis, Provost 2015)

Data, experience, rich simulations

Detect anomalies, unexpected variations, distributional shifts

Meta-analysis & transfer

Human engagement

http://dl.acm.org/citation.cfm?id=2700832

Predict previously unseen destination

Destinations, E1..,En, t

Day 1 Day 14

Predict new distinctions, combine open- & closed-world models

Krumm, H., 2006


http://research.microsoft.com/en-us/um/people/horvitz/predestination.pdf




Day 1 Day 14



Krumm, H., 2006





Day 1 Day 14



Krumm, H., 2006



2

1.5

x

1

0.5

00

0.5

y

11.5

20.2

0

-0.2

-0.4

-0.6

-0.8

-1

z

Direction: Joint modeling of key dimensions of error

Example: Learn about errors of perception & control

Probabilistic models of control

Probabilistic models of sensing 𝝋𝐨𝐛𝐬𝐭𝐚𝐜𝐥𝐞

𝝋𝐫𝐨𝐥𝐥

Sadigh & Kapoor, 2016

http://people.eecs.berkeley.edu/~dsadigh/Papers/sadigh-uncertainty-rss16.pdf

Proposed trajectory

Sample 1 Sample 2 Sample n

S.D.Mean

S.D.


𝝋𝐨𝐛𝐬𝐭𝐚𝐜𝐥𝐞




Proposed trajectory Trajectory safe if:

Σ

Σ Σ+> 1 − 𝜖

S.D.Mean

S.D.

Σ






Proposed trajectory

Sample 1 Sample 2 Sample n

Trajectory safe if:

Σ

Σ Σ+> 1 − 𝜖

S.D.Mean

S.D.

Σ



𝐩 > 𝟏 − 𝝐





Value of refining models & system- Value of additional data- Value of enhancing sensors- Value of better controller


(video)



𝐩 > 𝟏 − 𝝐


(video)



𝐩 > 𝟏 − 𝝐

Fail-safe


(video)



Sadigh & Kapoor, 2016 (video)


Direction: Verification, security, cryptography

Verification

Cryptography

Security

Safe AI Systems

Static analysis

Run-time verification

Whitebox fuzzing

Cybersecurity to protect attack surfaces

Appropriate use of physical security, isolation

Encryption for data integrity, protection of interprocess comms.

Direction: Verification, security, cryptography

Difficult to do formal analysis of large-scale system

Analysis & execution considers info. from running system

Direction: Runtime verification

Satisfy or violate desired properties?

Identify problem, future problem

Engage human

Take fail-safe action

State

Direction: Metalevel analysis, monitoring, assurance

EnvironmentAction

Environment’

State

AI system

State’

State


EnvironmentAction

Environment’

State

AI system

Reward

State’

Learning

State


EnvironmentAction

Environment’

State

AI system

Reward

State’

Learning

State


EnvironmentAction

Environment’

State

AI system

Reward

State’

Reinforcement

State

Learning

State


EnvironmentAction

Environment’

State

AI system

Reward Reinforcement

State’ Perception

Learning

State


EnvironmentAction

Environment’

State

AI system

Reward

State’ Perception

Reinforcement

Adversary

Learning

State


EnvironmentAction

Environment’

State

AI system

Reward

State’ Perception

Reinforcement

Adversary

Self-modification

e.g., see: Amodei, Olah, et al., 2016

Run-time verificationStatic analysis

Learning

State


EnvironmentAction

Environment’

State

AI system

Reward

State’ Perception

Reinforcement

Adversary

Reflective analysis• Operational faithfulness• Ensure isolation, detect mods• Identify external meddling

Direction: Human-machine collaboration

Models of human cognition

Transparency of state, explanation

Mastering coordination of initiatives


Air France 447 (June 2009)

China Airlines 006 (Feb 1985)

747 dives 10,000 in 20 seconds. 5g, supersonic.

Unrecoverable stall.


Air France 447 (June 2009)

China Airlines 006 (Feb 1985)

747 dives 10,000 in 20 seconds. 5g, supersonic.

Unrecoverable stall.

Machine learning & inference

Human cognition

Machineintelligence

How to best work together for safety?

Rich spectrum of autonomy


Kamar, Hacker, H., 2012

http://research.microsoft.com/en-us/um/people/horvitz/CrowdSynth.pdf

Designs for mix of initiatives

Machine learning & inference

Human cognition

Machineintelligence

How to best work together for safety?

Rich spectrum of autonomy


Kamar, Hacker, H., 2012

http://research.microsoft.com/en-us/um/people/horvitz/CrowdSynth.pdf



Infer challenges with machine competency

Infer challenges with machine competency


Infer human attention


p(a

tte

nti

on

sta

te|E

)

Time

Infer human attentionInfer challenges with machine competency


Continual prediction of trajectories

p(a

tte

nti

on

sta

te|E

)

Time




p(a

tte

nti

on

sta

te|E

)

Time




p(a

tte

nti

on

sta

te|E

)

Time



Safety-assuring mixed-initiative planner

- Driver’s attention over time

- Latency of human input

- Latency tolerance of situation

- Cost & influence of alerting driver

- Custom language, ongoing dialog

Gain driver attention t

Slow to defer need t’

Implement failsafe t’’

• Phases of study, testing, reporting for rolling out new capabilities in safety-critical domains (akin to FDA clinical trials, post-marketing surveillance)

• Disclosure & control of parameters on failure rates, tradeoffs, preferences

• Transparency & explainability of perception, inference, action

• System self-monitoring & reporting machinery

• Isolation of components in intelligence architectures

• Detecting & addressing feedback of system’s influence on self

Direction: Develop Best Practices for Safe AI

• Standard protocols for handoffs, attention, awareness, warning, in human-machine collaborations

• Policies for visible disclosure of autonomy to others (e.g., indication to others that a car is currently on automated policy)

• Fail-safe actions & procedures given predicted or sensed failures

• Enhancing robustness via co-design of environment & systems

• Testing for drift of assumptions, distributions in domains

• Special openness & adherence to best practices for data, learning, decision making for applications in governance & public policy

Direction: Develop Best Practices for Safe AI

Addressing concerns of public

Significant differences of opinion, including experts

Direction: Address concerns about “superintelligences”


Significant differences of opinion, including among experts


Alan Turing Script, BBC broadcast, 1951


Significant differences of opinion,


Alan Turing Script, BBC broadcast, 1951


Significant differences of opinion,


“For it seems possible that once the machine thinking method had started, it would not take long to outstrip our feeble powers.

…they would be able to converse with each other to sharpen their wits.

At some stage therefore, we should have to expect the machines to take control in the way that is mention in Samuel Butler’s Erewhon.”

Alan Turing, 1951

• Do we understand possibilities?

• What kind of research should done proactively?

• Can we “backcast” from imagined poor outcomes

• Designs of clear ways to thwart possibilities, ease concerns


Significant differences of opinion, including experts


Reflections on Safety and Artificial Intelligenceerichorvitz.com/OSTP-CMU_AI_Safety_framing_talk.pdfEric Horvitz Reflections on Safety and Artificial Intelligence Pittsburgh, PA June

Documents